Field-of-View Extension in Dental Cone-Beam CT via Implicit Neural Representations and Diffusion Model-Based Refinement
Susanne Schaub, Florentin Bieder, Matheus L. Oliveira, Yulan Wang, Buyanbileg Sodnom-ish, Dorothea Dagassan-Berndt, Michael M. Bornstein, Philippe C. Cattin
cs.CV · cs.AI
Quick verdict
Three-stage INR, iterative recon and diffusion pipeline extends dental CBCT FOV; strong simulated results, but no real-scan validation and GT circularity.
Summary
The paper tackles truncated field-of-view (FOV) reconstruction in dental cone-beam CT, where small detectors yield incomplete projection data. The authors propose a three-stage pipeline: an implicit neural representation (INR), extended with total-variation regularization, a normalization layer and hash-encoding regularization, overfits each truncated projection set and renders extrapolated projections; an OS-SART iterative reconstruction converts them into a 256-cubed volume; and a slice-wise conditional DDPM with only three denoising steps refines fine detail. On simulated projections from the public MMDental dataset (40 test scans), the full method reaches SSIM 0.960 and PSNR 41.38 dB versus 0.891 and 34.61 dB for OS-SART on cropped data, with ablations showing each stage contributes (INR-IR alone: SSIM 0.933; IR-DDPM: 0.955). A 1000-step DDPM scores slightly higher but needs about 2.2 h per volume, so the 3-step variant (about 25 s) is proposed. On ex vivo pig jaws (one validation and one test mandible), the method attains SSIM 0.985. Key limitations: all projections are simulated (TIGRE) and the ground truth is itself an OS-SART reconstruction of uncropped projections, so metrics partly measure consistency with that pipeline; generative hallucination outside the measured FOV is not quantified; the INR stage takes about 20 minutes; and no statistical significance testing is reported.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An INR trained per-scan on truncated projections supplies missing projection data well enough that subsequent OS-SART reconstruction substantially reduces truncation artifacts (INR-IR: SSIM 0.933 vs OS-SART 0.891 on MMDental).
- A slice-wise conditional DDPM with only 3 denoising steps recovers fine detail in ~25 s, approaching 1000-step quality (SSIM 0.960 vs 0.966 on MMDental) that would take ~2.2 h per volume.
- The full pipeline outperforms all tested baselines on both datasets (MMDental: SSIM 0.960 / PSNR 41.38 dB; pig jaws: SSIM 0.985 / PSNR 42.69 dB).
- The INR prior contributes more on pig jaws than on MMDental, where diffusion refinement alone (IR-DDPM, SSIM 0.955) nearly matches the full method (0.960).
Strengths
Component-isolating ablations
Section 2.5 defines INR-IR (no diffusion), IR-DDPM (no INR), IR-WDM and INR-IR-WDM (3D variants), plus a step-count sweep T=2/3/1000 (Table 1), so the contribution of each stage is traceable: e.g. MMDental INR-IR SSIM 0.933 -> full 0.960, and IR-DDPM 0.955 shows the DDPM alone is strong on MMDental but not on pig jaws (0.979 vs 0.985).
Large quantitative gains over truncated-data baselines
Table 1: proposed INR-IR-DDPM T=3 reaches SSIM 0.960±0.014 / PSNR 41.38±2.46 dB / LPIPS 0.100 on MMDental vs OS-SART 0.891 / 34.61 / 0.171; on pig jaws SSIM 0.985±0.006 / PSNR 42.69±2.14 dB.
Explicit speed/quality engineering
Two evaluation setups with public resources
Weaknesses
Circular ground truth
Simulation-only evaluation
Hallucination risk unquantified
Thin statistics and selective pig-jaw evaluation
Pipeline throughput bottleneck
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper tackles FOV extension in dental CBCT: reconstructing anatomy beyond a detector-truncated FOV from truncated projections. The three-stage pipeline (Section 2.2): (1) an INR (NAF-based, Section 2.1) with three added regularizers — TV (lambda=1e-5, Eq. 3), a normalization layer, and hash-encoding regularization — overfits each scan's truncated projections and renders extrapolated projections; (2) OS-SART reconstructs a 256-cubed volume (0.9 mm MMDental / 1.0 mm pig voxels); (3) a slice-wise conditional 2D DDPM with 3 denoising steps refines detail. Central claims: INR completion suppresses truncation artifacts (INR-IR SSIM 0.933 vs OS-SART 0.891); the fast DDPM nears 1000-step quality at 25 s vs 2.2 h; the full method beats all tested baselines (MMDental SSIM 0.960 / PSNR 41.38 dB; pig jaws SSIM 0.985 / PSNR 42.69 dB).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does this paper solve?
Dental cone-beam CT scanners with small detectors capture only part of the patient's head (a truncated field of view). The paper reconstructs anatomy beyond that truncated FOV from the incomplete projection data using a three-stage pipeline: a neural network (INR) estimates the missing projections, an iterative algorithm (OS-SART) rebuilds the 3D volume, and a fast diffusion model sharpens the result.
How well does the method perform?
On simulated projections from the public MMDental dataset (40 test scans), the full method reaches SSIM 0.960 and PSNR 41.38 dB, versus 0.891 and 34.61 dB for standard OS-SART on truncated data. On ex vivo pig jaws it reaches SSIM 0.985. The 3-step diffusion refinement takes about 25 seconds; the INR stage about 20 minutes per volume.
Was the method tested on real patient scans?
No. All projections were simulated with the TIGRE toolbox from existing volumes (the public MMDental dataset and ex vivo pig jaw scans). The authors themselves list validation on clinical scans with real projection data as future work, which is the main caveat for clinical use.
Is the code and data available?
The authors state the code is publicly available at github.com/SusanneSchaub/CBCT-FOV-Extension, and the main evaluation dataset (MMDental, 403 dental CBCT scans) is publicly available; the pig jaw scans come from a prior study by overlapping authors.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 2 (qualitative comparisons) cannot be assessed from extracted text; visual-quality and hallucination judgments rest on the authors' descriptions.
- The Table 1 runtime row is formatted oddly in the extraction; values were interpreted as WDM 3 s, DDPM 15 s/25 s/2.2 h at T=2/3/1000.
- Some protocol details (LPIPS backbone, SSIM 2D/3D aggregation, meaning of 'epochs' for per-scan INR overfitting) are not in the text and presumably live in the code repository.
- The evaluation setup description (ground truth as OS-SART recon of uncropped projections) is reconstructed from Section 3 text; a diagram of the data flow would disambiguate.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.